CernVM-FS on Supercomputers
There are several characteristics in which supercomputers can differ from other nodes with respect to CernVM-FS
Fuse is not allowed on the individual nodes
Individual nodes do not have Internet connectivity
Nodes have no local hard disk to store the CernVM-FS cache
These problems can be overcome as described in the following sections.
Running CernVM-FS as an unprivileged user
CernVM-FS can be run as an unprivileged user under several different scenarios. See documentation about that in the Security Running the client as a normal user section.
Parrot-Mounted CernVM-FS instead of Fuse Module
Instead of accessing
/cvmfs through a Fuse module, processes can use the
Parrot connector. The parrot
connector works on x86_64 Linux if the
ptrace system call is not disabled.
In contrast to a plain copy of a CernVM-FS repository to a shared file system,
this approach has the following advantages:
Millions of synchronized metadata operations per node (path lookups, in particular) will not drown the shared cluster file system but resolve locally in the parrot-cvmfs clients.
The file system is always consistent; applications never see half-synchronized directories.
After initial preloading, only change sets need to be transferred to the shared file system. This is much faster than rsync, which always has to browse the entire name space.
Identical files are internally deduplicated. While space of the order of terabytes is usually not an issue for HPC shared file systems, file system caches benefit from deduplication. It is also possible to preload only specific parts of a repository namespace.
Support for extra functionality implemented by CernVM-FS such as versioning and variant symlinks (symlinks resolved according to environment variables).
Downloading complete snapshots of CernVM-FS repositories
When there is no possible way to run the CernVM-FS client, an option that has been used on some HPC systems is to download entire or partial snapshots of CernVM-FS repositories using the cvmfs_shrinkwrap utility. These snapshots are also sometimes called “HPC fat container images”. This has many disadvantages compared to running a CernVM-FS client, so it is typically a last resort.
NFS Export with Cray DVS
Some HPC sites have tried running the cvmfs client on just one server and exporting to worker nodes over NFS. These installations can be made to work, but they are very inefficient, and often run into operational problems. If you want to try it out anyway using the Cray DVS please see the workaround on inode handling and DVS export.
NFS export is not a recommended setup to run cvmfs.
Preloading the CernVM-FS Cache
When the CernVM-FS client can be installed on the worker node but for
whatever reason on-demand downloading to a local cache is difficult, the
can be used to preload a CernVM-FS cache onto the shared cluster file system.
Internally it uses the same code that is used to replicate between CernVM-FS
stratum 0 and stratum 1. The
cvmfs_preload command is a self-extracting
binary with no further dependencies and should work on a majority of x86_64
Linux hosts. Note however that this method can significantly strain the
cluster file system’s metadata server(s) and that many HPC systems have
had better results with
as node caches as discussed below.
cvmfs_preload command replicates from a stratum 0 (not from a
stratum 1). Because this induces significant load on the source server,
stratum 0 administrators should be informed before using their server as a
source. As an example, in order to preload the ALICE repository into
/shared/cache, one could run from a login node
cvmfs_preload -u http://cvmfs-stratum-zero-hpc.cern.ch:8000/cvmfs/alice.cern.ch -r /shared/cache
This will preload the entire repository. In order to preload only specific parts of the namespace, you can create a _dirtab_ file with path prefixes. The path prefixes must not involve symbolic links. An example dirtab file for ALICE could look like
The corresponding invocation of
cvmfs_preload -u http://cvmfs-stratum-zero-hpc.cern.ch:8000/cvmfs/alice.cern.ch -r /shared/cache \
The initial preloading can take several hours to a few days. Subsequent invocations of the same command only transfer a change set and typically finish within seconds or minutes. These subsequent invocations need to be either done manually when necessary or scheduled for instance with a cron job.
cvmfs_preload command can preload files from multiple repositories
into the same cache directory.
Access from the Nodes
In order to access a preloaded cache from the nodes, set the path to the directory as an Alien Cache. Since there won’t be cache misses, parrot or fuse clients do not need to download additional files from the network.
If clients do have network access, they might find a repository version online
that is newer than the preloaded version in the cache. This results in
cvmfs_preload or in errors if the cache directory is
read-only. Therefore, we recommend to explicitly disable network access for the
parrot process on the nodes, for instance by setting
before the invocation of
cvmfs_preload from Sources
In order to compile
cvmfs_preload from sources, use the
-DBUILD_PRELOADER=yes cmake option.
Loopback File Systems for Nodes’ Caches
If nodes have Internet access but no local hard disk, it is preferable to provide the CernVM-FS caches as loopback file systems on the cluster file system. This way, CernVM-FS automatically populates the cache with the latest upstream content. A Fuse mounted CernVM-FS will also automatically manage the cache quota.
This approach requires a separate file for every node (not every mountpoint) on
the cluster file system. The file should be 15% larger than the configured
CernVM-FS cache size on the nodes, and it should be formatted with an ext3/4 or
an xfs file system. These files can be created with the
utilities. Nodes can mount these files as loopback file systems from the
shared file system.
Because there is only a single file for every node, the parallelism of the cluster file system can be exploited and all the requests from CernVM-FS circumvent the cluster file system’s metadata server(s). That can be a very large advantage because very often the metadata server is the bottleneck under typical workloads.